1
|
Graham MS, Tudosiu PD, Wright P, Pinaya WHL, Teikari P, Patel A, U-King-Im JM, Mah YH, Teo JT, Jäger HR, Werring D, Rees G, Nachev P, Ourselin S, Cardoso MJ. Latent Transformer Models for out-of-distribution detection. Med Image Anal 2023; 90:102967. [PMID: 37778102 PMCID: PMC10900071 DOI: 10.1016/j.media.2023.102967] [Citation(s) in RCA: 0] [Impact Index Per Article: 0] [Reference Citation Analysis] [Abstract] [Key Words] [MESH Headings] [Track Full Text] [Journal Information] [Subscribe] [Scholar Register] [Received: 12/08/2022] [Revised: 08/07/2023] [Accepted: 09/11/2023] [Indexed: 10/03/2023]
Abstract
Any clinically-deployed image-processing pipeline must be robust to the full range of inputs it may be presented with. One popular approach to this challenge is to develop predictive models that can provide a measure of their uncertainty. Another approach is to use generative modelling to quantify the likelihood of inputs. Inputs with a low enough likelihood are deemed to be out-of-distribution and are not presented to the downstream predictive model. In this work, we evaluate several approaches to segmentation with uncertainty for the task of segmenting bleeds in 3D CT of the head. We show that these models can fail catastrophically when operating in the far out-of-distribution domain, often providing predictions that are both highly confident and wrong. We propose to instead perform out-of-distribution detection using the Latent Transformer Model: a VQ-GAN is used to provide a highly compressed latent representation of the input volume, and a transformer is then used to estimate the likelihood of this compressed representation of the input. We demonstrate this approach can identify images that are both far- and near- out-of-distribution, as well as provide spatial maps that highlight the regions considered to be out-of-distribution. Furthermore, we find a strong relationship between an image's likelihood and the quality of a model's segmentation on it, demonstrating that this approach is viable for filtering out unsuitable images.
Collapse
Affiliation(s)
- Mark S Graham
- Department of Biomedical Engineering, School of Biomedical Engineering & Imaging Sciences, King's College London, London, UK.
| | - Petru-Daniel Tudosiu
- Department of Biomedical Engineering, School of Biomedical Engineering & Imaging Sciences, King's College London, London, UK
| | - Paul Wright
- Department of Biomedical Engineering, School of Biomedical Engineering & Imaging Sciences, King's College London, London, UK
| | - Walter Hugo Lopez Pinaya
- Department of Biomedical Engineering, School of Biomedical Engineering & Imaging Sciences, King's College London, London, UK
| | | | - Ashay Patel
- Department of Biomedical Engineering, School of Biomedical Engineering & Imaging Sciences, King's College London, London, UK
| | | | - Yee H Mah
- Department of Biomedical Engineering, School of Biomedical Engineering & Imaging Sciences, King's College London, London, UK; King's College Hospital NHS Foundation Trust, Denmark Hill, London, UK
| | - James T Teo
- King's College Hospital NHS Foundation Trust, Denmark Hill, London, UK; Institute of Psychiatry, Psychology & Neuroscience, King's College London, London, UK
| | - Hans Rolf Jäger
- Institute of Neurology, University College London, London, UK
| | - David Werring
- Stroke Research Centre, UCL Queen Square Institute of Neurology, London, UK
| | - Geraint Rees
- Institute of Neurology, University College London, London, UK
| | | | - Sebastien Ourselin
- Department of Biomedical Engineering, School of Biomedical Engineering & Imaging Sciences, King's College London, London, UK
| | - M Jorge Cardoso
- Department of Biomedical Engineering, School of Biomedical Engineering & Imaging Sciences, King's College London, London, UK
| |
Collapse
|
2
|
Neves P, McClure K, Verhoeven J, Dyubankova N, Nugmanov R, Gedich A, Menon S, Shi Z, Wegner JK. Global reactivity models are impactful in industrial synthesis applications. J Cheminform 2023; 15:20. [PMID: 36774523 PMCID: PMC9921076 DOI: 10.1186/s13321-023-00685-0] [Citation(s) in RCA: 3] [Impact Index Per Article: 3.0] [Reference Citation Analysis] [Abstract] [Track Full Text] [Figures] [Journal Information] [Subscribe] [Scholar Register] [Received: 11/16/2022] [Accepted: 01/22/2023] [Indexed: 02/13/2023] Open
Abstract
Artificial Intelligence is revolutionizing many aspects of the pharmaceutical industry. Deep learning models are now routinely applied to guide drug discovery projects leading to faster and improved findings, but there are still many tasks with enormous unrealized potential. One such task is the reaction yield prediction. Every year more than one fifth of all synthesis attempts result in product yields which are either zero or too low. This equates to chemical and human resources being spent on activities which ultimately do not progress the programs, leading to a triple loss when accounting for the cost of opportunity in time wasted. In this work we pre-train a BERT model on more than 16 million reactions from 4 different data sources, and fine tune it to achieve an uncertainty calibrated global yield prediction model. This model is an improvement upon state of the art not just from the increase in pre-train data but also by introducing a new embedding layer which solves a few limitations of SMILES and enables integration of additional information such as equivalents and molecule role into the reaction encoding, the model is called BERT Enriched Embedding (BEE). The model is benchmarked on an open-source dataset against a state-of-the-art synthesis focused BERT showing a near 20-point improvement in r2 score. The model is fine-tuned and tested on an internal company data benchmark, and a prospective study shows that the application of the model can reduce the total number of negative reactions (yield under 5%) ran in Janssen by at least 34%. Lastly, we corroborate the previous results through experimental validation, by directly deploying the model in an on-going drug discovery project and showing that it can also be used successfully as a reagent recommender due to its fast inference speed and reliable confidence estimation, a critical feature for industry application.
Collapse
Affiliation(s)
- Paulo Neves
- In-Silico Discovery and External Innovation (ISDEI), Janssen Research & Development, Janssen Pharmaceutica N.V, Beerse, Belgium.
| | - Kelly McClure
- Discovery Chemistry LJ, Janssen Research & Development, Janssen Pharmaceutica N.V, Philadelphia, United States of America
| | - Jonas Verhoeven
- In-Silico Discovery and External Innovation (ISDEI), Janssen Research & Development, Janssen Pharmaceutica N.V, Beerse, Belgium
| | - Natalia Dyubankova
- In-Silico Discovery and External Innovation (ISDEI), Janssen Research & Development, Janssen Pharmaceutica N.V, Beerse, Belgium
| | - Ramil Nugmanov
- In-Silico Discovery and External Innovation (ISDEI), Janssen Research & Development, Janssen Pharmaceutica N.V, Beerse, Belgium
| | | | - Sairam Menon
- Pharma R&D Information Tech, Janssen Research & Development, Janssen Pharmaceutica N.V, Beerse, Belgium
| | - Zhicai Shi
- Discovery Chemistry LJ, Janssen Research & Development, Janssen Pharmaceutica N.V, Philadelphia, United States of America
| | - Jörg K Wegner
- In-Silico Discovery and External Innovation (ISDEI), Janssen Research & Development, Janssen Pharmaceutica N.V, Beerse, Belgium
| |
Collapse
|
3
|
Can uncertainty estimation predict segmentation performance in ultrasound bone imaging? Int J Comput Assist Radiol Surg 2022; 17:825-832. [PMID: 35377036 DOI: 10.1007/s11548-022-02597-0] [Citation(s) in RCA: 3] [Impact Index Per Article: 1.5] [Reference Citation Analysis] [Abstract] [Key Words] [Track Full Text] [Journal Information] [Subscribe] [Scholar Register] [Received: 03/03/2022] [Accepted: 03/08/2022] [Indexed: 11/05/2022]
Abstract
PURPOSE Segmenting bone surfaces in ultrasound (US) is a fundamental step in US-based computer-assisted orthopaedic surgeries. Neural network-based segmentation techniques are a natural choice for this, given promising results in related tasks. However, to gain widespread use, we must be able to know how much to trust segmentation networks during clinical deployment when ground-truth data is unavailable. METHODS We investigated alternative ways to measure the uncertainty of trained networks by implementing a baseline U-Net trained on a large dataset, together with three uncertainty estimation modifications: Monte Carlo dropout, test time augmentation, and ensemble learning. We measured the segmentation performance, calibration quality, and the ability to predict segmentation performance on test data. We further investigated the effect of data quality on these measures. RESULTS Overall, we found that ensemble learning with binary cross-entropy (BCE) loss achieved the best segmentation performance (mean Dice: 0.75-0.78 and RMS distance: 0.62-0.86mm) and the lowest calibration errors (mean: 0.22-0.28%). In contrast to previous studies of area or volumetric segmentation, we found that the resulting uncertainty measures are not reliable proxies for surface segmentation performance. CONCLUSION Our experiments indicate that a significant performance and confidence calibration boost can be achieved with ensemble learning and BCE loss, as tested on 13,687 US images containing various anatomies and imaging parameters. However, these techniques do not allow us to reliably predict future segmentation performance. The results of this study can be used to improve the calibration and performance of US segmentation networks.
Collapse
|
4
|
Monte Carlo Dropout for Uncertainty Estimation and Motor Imagery Classification. SENSORS 2021; 21:s21217241. [PMID: 34770553 PMCID: PMC8588128 DOI: 10.3390/s21217241] [Citation(s) in RCA: 0] [Impact Index Per Article: 0] [Reference Citation Analysis] [Abstract] [Key Words] [Track Full Text] [Download PDF] [Figures] [Subscribe] [Scholar Register] [Received: 09/24/2021] [Revised: 10/24/2021] [Accepted: 10/28/2021] [Indexed: 11/16/2022]
Abstract
Motor Imagery (MI)-based Brain-Computer Interfaces (BCIs) have been widely used as an alternative communication channel to patients with severe motor disabilities, achieving high classification accuracy through machine learning techniques. Recently, deep learning techniques have spotlighted the state-of-the-art of MI-based BCIs. These techniques still lack strategies to quantify predictive uncertainty and may produce overconfident predictions. In this work, methods to enhance the performance of existing MI-based BCIs are proposed in order to obtain a more reliable system for real application scenarios. First, the Monte Carlo dropout (MCD) method is proposed on MI deep neural models to improve classification and provide uncertainty estimation. This approach was implemented using Shallow Convolutional Neural Network (SCNN-MCD) and with an ensemble model (E-SCNN-MCD). As another contribution, to discriminate MI task predictions of high uncertainty, a threshold approach is introduced and tested for both SCNN-MCD and E-SCNN-MCD approaches. The BCI Competition IV Databases 2a and 2b were used to evaluate the proposed methods for both subject-specific and non-subject-specific strategies, obtaining encouraging results for MI recognition.
Collapse
|
5
|
Bayesian U-Net: Estimating Uncertainty in Semantic Segmentation of Earth Observation Images. REMOTE SENSING 2021. [DOI: 10.3390/rs13193836] [Citation(s) in RCA: 6] [Impact Index Per Article: 2.0] [Reference Citation Analysis] [Abstract] [Track Full Text] [Subscribe] [Scholar Register] [Indexed: 11/16/2022]
Abstract
In recent years, numerous deep learning techniques have been proposed to tackle the semantic segmentation of aerial and satellite images, increase trust in the leaderboards of main scientific contests and represent the current state-of-the-art. Nevertheless, despite their promising results, these state-of-the-art techniques are still unable to provide results with the level of accuracy sought in real applications, i.e., in operational settings. Thus, it is mandatory to qualify these segmentation results and estimate the uncertainty brought about by a deep network. In this work, we address uncertainty estimations in semantic segmentation. To do this, we relied on a Bayesian deep learning method, based on Monte Carlo Dropout, which allows us to derive uncertainty metrics along with the semantic segmentation. Built on the most widespread U-Net architecture, our model achieves semantic segmentation with high accuracy on several state-of-the-art datasets. More importantly, uncertainty maps are also derived from our model. While they allow for the performance of a sounder qualitative evaluation of the segmentation results, they also include valuable information to improve the reference databases.
Collapse
|
6
|
A Novel Hybrid Approach Based on Deep CNN Features to Detect Knee Osteoarthritis. SENSORS 2021; 21:s21186189. [PMID: 34577402 PMCID: PMC8471198 DOI: 10.3390/s21186189] [Citation(s) in RCA: 11] [Impact Index Per Article: 3.7] [Reference Citation Analysis] [Abstract] [Key Words] [Track Full Text] [Download PDF] [Figures] [Subscribe] [Scholar Register] [Received: 06/23/2021] [Revised: 09/09/2021] [Accepted: 09/10/2021] [Indexed: 11/17/2022]
Abstract
In the recent era, various diseases have severely affected the lifestyle of individuals, especially adults. Among these, bone diseases, including Knee Osteoarthritis (KOA), have a great impact on quality of life. KOA is a knee joint problem mainly produced due to decreased Articular Cartilage between femur and tibia bones, producing severe joint pain, effusion, joint movement constraints and gait anomalies. To address these issues, this study presents a novel KOA detection at early stages using deep learning-based feature extraction and classification. Firstly, the input X-ray images are preprocessed, and then the Region of Interest (ROI) is extracted through segmentation. Secondly, features are extracted from preprocessed X-ray images containing knee joint space width using hybrid feature descriptors such as Convolutional Neural Network (CNN) through Local Binary Patterns (LBP) and CNN using Histogram of oriented gradient (HOG). Low-level features are computed by HOG, while texture features are computed employing the LBP descriptor. Lastly, multi-class classifiers, that is, Support Vector Machine (SVM), Random Forest (RF), and K-Nearest Neighbour (KNN), are used for the classification of KOA according to the Kellgren-Lawrence (KL) system. The Kellgren-Lawrence system consists of Grade I, Grade II, Grade III, and Grade IV. Experimental evaluation is performed on various combinations of the proposed framework. The experimental results show that the HOG features descriptor provides approximately 97% accuracy for the early detection and classification of KOA for all four grades of KL.
Collapse
|
7
|
Abstract
SAR image registration is a crucial problem in SAR image processing since the registration results with high precision are conducive to improving the quality of other problems, such as change detection of SAR images. Recently, for most DL-based SAR image registration methods, the problem of SAR image registration has been regarded as a binary classification problem with matching and non-matching categories to construct the training model, where a fixed scale is generally set to capture pair image blocks corresponding to key points to generate the training set, whereas it is known that image blocks with different scales contain different information, which affects the performance of registration. Moreover, the number of key points is not enough to generate a mass of class-balance training samples. Hence, we proposed a new method of SAR image registration that meanwhile utilizes the information of multiple scales to construct the matching models. Specifically, considering that the number of training samples is small, deep forest was employed to train multiple matching models. Moreover, a multi-scale fusion strategy is proposed to integrate the multiple predictions and obtain the best pair matching points between the reference image and the sensed image. Finally, experimental results on four datasets illustrate that the proposed method is better than the compared state-of-the-art methods, and the analyses for different scales also indicate that the fusion of multiple scales is more effective and more robust for SAR image registration than one single fixed scale.
Collapse
|